Mercurial > ift6266

comparison deep/stacked_dae/v_sylvain/nist_sda.py @ 238:9fc641d7adda

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Possibilite de restreindre la taille des ensemble d'entrainement, valid et test afin de pouvoir tester le code rapidement
author SylvainPL <sylvain.pannetier.lebeuf@umontreal.ca>
date Mon, 15 Mar 2010 13:22:20 -0400
parents ecb69e17950b
children 6d49cf134a40
comparison
equal deleted inserted replaced
237:9b6e0af062af 238:9fc641d7adda
47 47
48 if TEST_CONFIG: 48 if TEST_CONFIG:
49 REDUCE_TRAIN_TO = 1000 49 REDUCE_TRAIN_TO = 1000
50 MAX_FINETUNING_EPOCHS = 2 50 MAX_FINETUNING_EPOCHS = 2
51 REDUCE_EVERY = 10 51 REDUCE_EVERY = 10
52 MINIBATCH_SIZE=20
52 53
53 # Possible values the hyperparameters can take. These are then 54 # Possible values the hyperparameters can take. These are then
54 # combined with produit_cartesien_jobs so we get a list of all 55 # combined with produit_cartesien_jobs so we get a list of all
55 # possible combinations, each one resulting in a job inserted 56 # possible combinations, each one resulting in a job inserted
56 # in the jobman DB. 57 # in the jobman DB.
69 'pretraining_epochs_per_layer':2, 70 'pretraining_epochs_per_layer':2,
70 'max_finetuning_epochs':2, 71 'max_finetuning_epochs':2,
71 'hidden_layers_sizes':500, 72 'hidden_layers_sizes':500,
72 'corruption_levels':0.2, 73 'corruption_levels':0.2,
73 'minibatch_size':20, 74 'minibatch_size':20,
74 #'reduce_train_to':10000, 75 'reduce_train_to':10000,
75 'num_hidden_layers':1}) 76 'num_hidden_layers':1})
76 77
77 ''' 78 '''
78 Function called by jobman upon launching each job 79 Function called by jobman upon launching each job
79 Its path is the one given when inserting jobs: see EXPERIMENT_PATH 80 Its path is the one given when inserting jobs: see EXPERIMENT_PATH
92 ## 93 ##
93 ## nist = NIST(minibatch_size=20) 94 ## nist = NIST(minibatch_size=20)
94 ## 95 ##
95 ## print "NIST loaded" 96 ## print "NIST loaded"
96 ## 97 ##
97 ## # For test runs, we don't want to use the whole dataset so 98 # For test runs, we don't want to use the whole dataset so
98 ## # reduce it to fewer elements if asked to. 99 # reduce it to fewer elements if asked to.
99 ## rtt = None 100 rtt = None
100 ## if state.has_key('reduce_train_to'): 101 if state.has_key('reduce_train_to'):
101 ## rtt = state['reduce_train_to'] 102 rtt = int(state['reduce_train_to']/state['minibatch_size'])
102 ## elif REDUCE_TRAIN_TO: 103 elif REDUCE_TRAIN_TO:
103 ## rtt = REDUCE_TRAIN_TO 104 rtt = int(REDUCE_TRAIN_TO/MINIBATCH_SIZE)
104 ## 105
105 ## if rtt: 106 if rtt:
106 ## print "Reducing training set to "+str(rtt)+ " examples" 107 print "Reducing training set to "+str(rtt*state['minibatch_size'])+ " examples"
108 else:
109 rtt=float('inf') #No reduction
107 ## nist.reduce_train_set(rtt) 110 ## nist.reduce_train_set(rtt)
108 ## 111 ##
109 ## train,valid,test = nist.get_tvt() 112 ## train,valid,test = nist.get_tvt()
110 ## dataset = (train,valid,test) 113 ## dataset = (train,valid,test)
111 114
112 n_ins = 32*32 115 n_ins = 32*32
113 n_outs = 62 # 10 digits, 26*2 (lower, capitals) 116 n_outs = 62 # 10 digits, 26*2 (lower, capitals)
114 117
115 series = create_series(state.num_hidden_layers) 118 series = create_series(state.num_hidden_layers)
116 119
117 print "Creating optimizer with state, ", state 120 print "Creating optimizer with state, ", state
118 121
119 optimizer = SdaSgdOptimizer(dataset=datasets.nist_all, hyperparameters=state, \ 122 optimizer = SdaSgdOptimizer(dataset=datasets.nist_all, hyperparameters=state, \
120 n_ins=n_ins, n_outs=n_outs,\ 123 n_ins=n_ins, n_outs=n_outs,\
121 series=series) 124 series=series)
122 125
123 optimizer.pretrain(datasets.nist_all) 126 optimizer.pretrain(datasets.nist_all,rtt)
124 channel.save() 127 channel.save()
125 128
126 optimizer.finetune(datasets.nist_all) 129 optimizer.finetune(datasets.nist_all,rtt)
127 channel.save() 130 channel.save()
128 131
129 return channel.COMPLETE 132 return channel.COMPLETE
130 133
131 # These Series objects are used to save various statistics 134 # These Series objects are used to save various statistics